Classification of EEG Signals on SEED Dataset Using Improved CNN

Abstract

The proposed research introduces an Improved Convolutional Neural Network (ICNN) to construct EEG-based emotion detection models. This study has utilized an EEG dataset of 15 subjects available from a BCMI laboratory. In our work, differential entropy characteristics obtained from multichannel EEG data are used to train the Improved CNN. The best classification accuracy is 95.67% which is significantly higher than that of the original 62 channels. The most important channels and frequency bands are identified by Improved CNN. The outcomes of our study also demonstrate the existence of neuronal signatures linked to various emotions, which are consistent between sessions and people. Finally, the effectiveness of deep and shallow models are compared and also the performance of improved CNN is compared with benchmark algorithms.